fix: count malformed surrogate pairs as separate U+FFFD replacements in the repair cap

The 4-byte branch of utf8BytesExceed skipped the next code unit whenever it
existed, without checking it is a low surrogate — so a run of high surrogates
was counted at 2 bytes per unit while TextEncoder emits 3. The pair path now
requires the next unit in 0xDC00..0xDFFF; anything else counts 3 bytes without
skipping. Regression: 200k high-surrogate pairs (400k code units, 1.2 MB on
the wire) decline repair instead of being admitted at 800k counted bytes.

Author: lidge-jun

src/adapters/anthropic.ts

@@ -299,7 +299,11 @@ function utf8BytesExceed(input: string, max: number): boolean {
     const code = input.charCodeAt(i);
     if (code < 0x80) bytes += 1;
     else if (code < 0x800) bytes += 2;
-    else if (code >= 0xd800 && code <= 0xdbff && i + 1 < input.length) {
+    else if (code >= 0xd800 && code <= 0xdbff && i + 1 < input.length
+      && input.charCodeAt(i + 1) >= 0xdc00 && input.charCodeAt(i + 1) <= 0xdfff) {
+      // A complete surrogate pair is one 4-byte scalar. Anything else — a high surrogate
+      // followed by another high surrogate or a non-surrogate — encodes as two separate
+      // U+FFFD replacements, so the next unit must NOT be skipped.
       bytes += 4;
       i++;
     } else bytes += 3; // lone surrogates encode as U+FFFD (3 bytes)

tests/anthropic-eof-tolerance.test.ts

@@ -190,4 +190,16 @@ describe("AgentRouter Anthropic EOF tolerance (#658)", () => {
     const events = await createAnthropicAdapter(tolerant).parseResponse(new Response(payload));
     expect(events).toContainEqual({ type: "tool_call_delta", arguments: "{}" });
   });
+
+  test("malformed surrogate pairs count as separate U+FFFD replacements against the cap", async () => {
+    // 200k high-surrogate pairs: 400k code units, but every lone surrogate encodes as its
+    // own 3-byte U+FFFD — 1.2 MB over the wire, which a pair-skipping count would admit.
+    const oversized = `${"\ud800\ud800".repeat(200_000)}{"ok":true}`;
+    const payload = JSON.stringify({
+      content: [{ type: "tool_use", id: "toolu_1", name: "get_weather", input: oversized }],
+    });
+
+    const events = await createAnthropicAdapter(tolerant).parseResponse(new Response(payload));
+    expect(events).toContainEqual({ type: "tool_call_delta", arguments: "{}" });
+  });
 });